Google paper: Self-Improvement Memory lifts cross-model harness transfer from 68.41 to 84.75
rohanpaul_ai · x · 2026-11-18
A Google/Penn State paper introduces Continual Recursive Self-Improvement (CRSI) and Self-Improvement Memory (SIM) for agent harnesses — the code that runs prompts and tools.
- SIM is maintained by the harness optimizer: each edit is logged with the problem it targeted, reasoning, and measured outcomes
- When the harness moves to a new model or benchmark, the optimizer reads the log to decide which edits to keep or revise
- Results: an Opus-evolved harness moved to Gemini 3.8 Flash workspace tasks scored 84.75 with SIM vs 73.42 without (from a 68.41 baseline)
- Across 10 coding runs, SIM finished ahead 9 times, by an average of 2.36 points (p=0.003)
Paper: arXiv:2611.18856, covering coding, text classification, workspace and engineering domains.
Related event: Google Proposes Self-Improvement Memory for Cross-Model Agent Harnesses(3 posts)→
More from coding & agent
- Cognition launches Devin Bots: one bot that leads a team of agents to ship work — round · 2026-11-18
- Shared applications, not shared chat: a new blueprint for collaborating with agents — thechrisperry · 2026-11-18
- Microsoft publishes playbook for enterprise deployment of governed AI agents on Windows via MXC — pavandavuluri · 2026-11-18
- AWS tests three OAuth patterns for agents calling Snowflake Cortex via MCP — AWS ML Blog · 2026-11-18
- A16z partner Anish Acharya curates a list of the sharpest thinkers on agent swarms — astrange · 2026-11-18
- Netlify CLI 27.18.0 moves auth tokens from plain text into OS keychains — jasonkneen · 2026-11-18